高動態範圍 (HDR) 成像需要融合在同一場景中以多種不同曝光程度的影像以覆蓋整個動態範圍。以目前現有的研究中，只利用少數低動態範圍 (LDR) 影像，這仍然是一項具有挑戰性的任務。本論文提出了一種新穎的兩曝光影像融合模型，此模型具有我們提出的交叉注意力融合模組 (CAFM)，可使用一個影像的高曝光的部分來補償因曝光不足或過度曝光而導致的另一張影像內容缺失的部分。CAFM 由 交叉注意力融合(Cross Attention Fusion) 和 通道注意力融合(Channel Attention Fusion) 組成，以實現雙分支融合，從而產生出色的融合結果。並且在公開的HDR 資料集上，我們進行大量實驗以證明所提出的模型在與最先驅的圖像融合方法比較時表現良好。

High Dynamic Range (HDR) imaging requires the fusion of images captured with multiple exposure ratios in the same scene to cover the entire dynamic range. With only a few low dynamic range (LDR) images, it remains a challenging task. The paper presents a novel two-exposure image fusion model that features the proposed Cross Attention Fusion Module (CAFM) to use one image's highlight to compensate for the other's content loss caused by under-exposure or over-exposure. The CAFM consists of Cross Attention Fusion and Channel Attention Fusion to achieve a dual-branch fusion for producing superior fusion results. The extensive experimental results on benchmark HDR public datasets demonstrate that the proposed model performs favorably against the state-of-the-art image fusion methods.

[1] C. Florea, C. Vertan, and L. Florea, “High dynamic range imaging by perceptuallogarithmic exposure merging,”International Journal of Applied Mathematics andComputer Science, vol. 25, no. 4, pp. 943–954, 2015.
[2] T. Mertens, J. Kautz, and F. Van Reeth, “Exposure fusion: A simple and practical al­ternativetohighdynamicrangephotography,”inComputer graphics forum,vol.28,pp. 161–171, Wiley Online Library, 2009.
[3] F. Kou, Z. Li, C. Wen, and W. Chen, “Multi­scale exposure fusion via gradient do­mainguidedimagefiltering,”in2017 IEEE International Conference on Multimediaand Expo (ICME), pp. 1105–1110, IEEE, 2017.
[4] Y. Yang, W. Cao, S. Wu, and Z. Li, “Multi­scale fusion of two large­exposure­ratioimages,” 2018.
[5] K. R. Prabhakar, V. S. Srikar, and R. V. Babu, “Deepfuse: A deep unsupervisedapproach for exposure fusion with extreme exposure image pairs.,” inICCV, vol. 1,p. 3, 2017.
[6] G. Eilertsen, J. Kronander, G. Denes, R. K. Mantiuk, and J. Unger, “Hdr image re­constructionfromasingleexposureusingdeepcnns,”ACM transactions on graphics(TOG), vol. 36, no. 6, pp. 1–15, 2017.
[7] Y. Endo, Y. Kanamori, and J. Mitani, “Deep reverse tone mapping.,”ACM Trans.Graph., vol. 36, no. 6, pp. 177–1, 2017.
[8] Y. Chen, M. Yu, K. Chen, G. Jiang, Y. Song, Z. Peng, and F. Chen, “New stereo high dynamic range imaging method using generative adversarial networks,”in2019IEEE International Conference on Image Processing (ICIP), pp. 3502–3506, IEEE,2019.
[9] J.­L. Yin, B.­H. Chen, Y.­T. Peng, and C.­C. Tsai, “Deep prior guided network forhigh­quality image fusion,” in2020 IEEE International Conference on Multimediaand Expo (ICME), pp. 1–6, IEEE, 2020.32
[10] H. Xu, J. Ma, Z. Le, J. Jiang, and X. Guo, “Fusiondn: A unified densely connectednetworkforimagefusion,”inProceedings of the Thirty­Fourth AAAI Conference onArtificial Intelligence (AAAI), pp. 12484–12491, 2020.
[11] J. Hu, L. Shen, and G. Sun, “Squeeze­and­excitation networks,” inProceedings ofthe IEEE conference on computer vision and pattern recognition, pp. 7132–7141,2018.
[12] S. Woo, J. Park, J.­Y. Lee, and I. So Kweon, “Cbam: Convolutional block attentionmodule,” inProceedings of the European conference on computer vision (ECCV),pp. 3–19, 2018.
[13] X. Li, W. Wang, X. Hu, and J. Yang, “Selective kernel networks,” inProceedingsof the IEEE conference on computer vision and pattern recognition, pp. 510–519,2019.
[14] R. Qian, R. T. Tan, W. Yang, J. Su, and J. Liu, “Attentive generative adversarialnetwork for raindrop removal from a single image,” inProceedings of the IEEEconference on computer vision and pattern recognition, pp. 2482–2491, 2018.
[15] F. Lv, Y. Li, and F. Lu, “Attention guided low­light image enhancement with a largescale low­light simulation dataset,”arXiv: 1908.00682, 2019.
[16] J. Fu, J. Liu, H. Tian, Y. Li, Y. Bao, Z. Fang, and H. Lu, “Dual attention networkfor scene segmentation,” inProceedings of the IEEE/CVF Conference on ComputerVision and Pattern Recognition, pp. 3146–3154, 2019.
[17] Q. Hou, L. Zhang, M.­M. Cheng, and J. Feng, “Strip Pooling: Rethinking spatialpooling for scene parsing,” 2020.
[18] H. Yeganeh and Z. Wang, “Objective quality assessment of tone­mapped images,”IEEE Transactions on Image Processing, vol. 22, no. 2, pp. 657–667, 2012.
[19] K. Gu, S. Wang, G. Zhai, S. Ma, X. Yang, W. Lin, W. Zhang, and W. Gao, “Blindquality assessment of tone­mapped images via analysis of information, naturalness,and structure,”IEEE Transactions on Multimedia, 2016.
[20] J. Cai, S. Gu, and L. Zhang, “Learning a deep single image contrast enhancer frommulti­exposure images,”IEEE Transactions on Image Processing, vol. 27, no. 4,pp. 2049–2062, 2018.33
[21] Q. Wang, W. Chen, X. Wu, and Z. Li, “Detail­enhanced multi­scale exposure fusionin yuv color space,” 2019.
[22] M. Nejati, M. Karimi, S. R. Soroushmehr, N. Karimi, S. Samavi, and K. Najar­ian, “Fast exposure fusion using exposedness function,” in2017 IEEE InternationalConference on Image Processing (ICIP), pp. 2234–2238, IEEE, 2017.
[23] K. Ma, H. Li, H. Yong, Z. Wang, D. Meng, and L. Zhang, “Robust multi­exposureimage fusion: a structural patch decomposition approach,”IEEE Transactions onImage Processing, vol. 26, no. 5, pp. 2519–2532, 2017.
[24] A. Rafi, M. Tinauli, and M. Izani, “High dynamic range images: Evolution, applica­tions and suggested processes,” in2007 11th International Conference InformationVisualization (IV’07), pp. 877–882, IEEE, 2007.
[25] Y. Kinoshita and H. Kiya, “Scene segmentation­based luminance adjustment formulti­exposure image fusion,”IEEE Transactions on Image Processing, vol. 28,no. 8, pp. 4101–4116, 2019.
[26] Y. Kinoshita, T. Yoshida, S. Shiota, and H. Kiya, “Pseudo multi­exposure fusionusing a single image,” in2017 Asia­Pacific Signal and Information Processing As­sociation Annual Summit and Conference (APSIPA ASC), pp. 263–269, IEEE, 2017.
[27] Y. Kinoshita and H. Kiya, “Automatic exposure compensation using an image seg­mentationmethodforsingle­image­basedmulti­exposurefusion,”APSIPA Transac­tions on Signal and Information Processing, vol. 7, 2018.
[28] A. Visavakitcharoen, Y. Kinoshita, and H. Kiya, “A color compensation methodusing inverse camera response function for multi­exposure image fusion,” in2019IEEE 8th Global Conference on Consumer Electronics (GCCE),pp.468–470,IEEE,2019.
[29] Z.Li, Z.Wei, C.Wen, andJ.Zheng, “Detail­enhancedmulti­scaleexposurefusion,”IEEE Transactions on Image processing, vol. 26, no. 3, pp. 1243–1252, 2017.
[30] T. Sakai, D. Kimura, T. Yoshida, and M. Iwahashi, “Hybrid method for multi­exposure image fusion based on weighted mean and sparse representation,” in201523rd European Signal Processing Conference (EUSIPCO), pp. 809–813, IEEE,2015.34
[31] N.K.KalantariandR.Ramamoorthi,“Deephighdynamicrangeimagingofdynamicscenes.,”ACM Trans. Graph., vol. 36, no. 4, pp. 144–1, 2017.
[32] S. Wu, J. Xu, Y.­W. Tai, and C.­K. Tang, “Deep high dynamic range imaging withlargeforegroundmotions,”inProceedings of the European Conference on ComputerVision (ECCV), pp. 117–132, 2018.
[33] K. Ma, K. Zeng, and Z. Wang, “Perceptual quality assessment for multi­exposureimage fusion,”IEEE Transactions on Image Processing, vol. 24, no. 11, pp. 3345–3356, 2015.
[34] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connectedconvolutionalnetworks,” inProceedings of the IEEE conference on computer visionand pattern recognition, pp. 4700–4708, 2017.
[35] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser,and I. Polosukhin, “Attention is all you need,” inAdvances in neural informationprocessing systems, pp. 5998–6008, 2017.
[36] Z.Pu,P.Guo,M.S.Asif,andZ.Ma,“Robusthighdynamicrange(hdr)imagingwithcomplexmotionandparallax,”inProceedings of the Asian Conference on ComputerVision, 2020.